Formic Robotics

The Real World Rewards Explorers, Not Memorizers In the previous post, we explained why robots don’t need bigger brains (or bigger LLMs). Now let’s put that claim to the test with a simple - but surprisingly tough - example: plugging in a wobbly USB-C cable. At first glance, it’s trivial for a human. For AI, it’s a nightmare. Here’s why even bigger LLMs still struggle: Words ≠ actions: Predicting text isn’t the same as controlling motors at high frequency. Exploding complexity: Tiny shifts in angle create massive uncertainty. Skills simplify it: align → probe → insert. Partial observability: The cable tip blocks the view - only micro-probes reveal what’s actually happening. Rare surprises: Bent pins or sticky fits can’t be memorized. They demand adaptation, in real time. Safety first: Robots need force limits and aborts, not just next-token guesses. LLMs still play a role - they can draft task graphs, suggest parameters, or summarize telemetry. But the core competence isn’t language. It’s: a library of skills, plus a planner that knows when to use them - and when to explore. Curiosity > Memorization Memorization assumes tomorrow looks like yesterday. Curiosity, by contrast, asks: Are we aligned? Is friction high? Did the object slip? Small, safe experiments - taps, wiggles, angle sweeps - update beliefs and unlock robust behavior even when the world is messy. The Compounding Win: Incremental Growth Start with a few reliable skills and simple plans. Collect targeted real-world data. Improve the weakest skill (no need to retrain the whole system). Add new skills - the planner simply learns when to call them. That’s agentic AI for embodiment: focused, modular, and improving week by week. So - can you think of other “USB moments” where AI breaks down? Do you agree that in robotics (and maybe in life), curiosity beats memorization? #Robotics #ArtificialIntelligence #AgenticAI #MachineLearning #EmbodiedAI #ReinforcementLearning #FutureOfAI #Automation

Robots Don’t Need Bigger Brains  -  They Need Better Skills (and a Bit of Curiosity) Think about how you learned to drive. You didn’t memorize the internet of driving - every trip ever taken in human history, like large language models (LLMs) do... Instead, you built up bounded skills: - Lane keeping - Mirror checks - Parallel parking Each skill was compact: clear inputs, a few outputs, and obvious success criteria. Then you learned how to stitch them together with simple rules: - Signal - Check blind spot - Merge into lane And you stayed curious. You probed new situations - rain, hills, night driving- not by memorizing them in advance, but by carefully testing how the car responded and adapting your skills. That combination of task-specific skills, lightweight orchestration, and safe exploration is exactly what robots need. What does this teach us? Large language models are extraordinary at language. But the physical world isn’t a neat paragraph. It’s noisy, continuous, and full of odd edge cases. Scaling up data and compute helps, but when it comes to embodiment - getting machines to act in the real world - we need a different center of gravity: - Skills that shrink complex problems into tractable pieces - Orchestration that connects those skills into larger behaviors - Curiosity that drives safe exploration when data runs out Robots don’t need ever-bigger brains. They need the right skills and curiosity, organized to grow steadily more capable in the messy reality we live in. Or in technical terms: Agentic AI, composed of skills, sequencing and curiosity. What do you think? #AgenticAI #ArtificialIntelligence #DeepLearning #RobotAutonomy #EmbodiedAI #CognitiveAI #AutonomousSystems #FutureOfAI

We’re #hiring a new Senior SLAM Engineer in Tel Aviv District. Apply today or share this post with your network.

The future belongs to the rapidly adaptive #Tech #Robotics #AI #Unbundle #AgenticAI #PhysicalAI

"Everyone is going to be augmented by AI"... #AI #AgenticAI #PhysicalAI #Robotics